24 Jun
2006
24 Jun
'06
1:15 p.m.
Hi!
Trying to understand some magic with Ipmulse Response, FFT and so on, I'm playing
with
scipy, matplotlib (and python as new language for me as a bonus). Concrete IR was got
with help of DRC (rec_imp, saying more strictly).
Currently I have got this frequency response graph:
http://gaydenko.com/mix/needSmooth.png
And the question is: what is common way to smooth the result? Some offtopic apps has
something like "1/24, ..., 1/3 octave smoothing". What does it mean?
Andrew
24 Jun
24 Jun
1:29 p.m.
On Sat, Jun 24, 2006 at 07:15:38PM +0400, Andrew Gaydenko wrote:
And the question is: what is common way to smooth the
result? Some offtopic apps has
something like "1/24, ..., 1/3 octave smoothing". What does it mean?
Lowpass (linear phase) filtering of the response. Or windowing the
IR which is equivalent.
If you window the IR, you get a constant resolution over the
entire range of the FR. To get a variable rate such as 1/3 oct,
you need frequency dependent windowing - the window gets shorter
as frequency rises. It's not easy to do right (one of the reasons
why aliki still isn't finished).
IIRC DRC can produce such plots if you ask it nicely.
--
FA
Follie! Follie! Delirio vano e' questo!
2:04 p.m.
======= On Saturday 24 June 2006 19:30, Fons Adriaensen wrote: =======
.....
Lowpass (linear phase) filtering of the response. Or windowing the
IR which is equivalent.
If you window the IR, you get a constant resolution over the
entire range of the FR. To get a variable rate such as 1/3 oct,
you need frequency dependent windowing - the window gets shorter
as frequency rises. It's not easy to do right (one of the reasons
why aliki still isn't finished).
IIRC DRC can produce such plots if you ask it nicely.
--
FA
Follie! Follie! Delirio vano e' questo!
========================================================================
Hi, Fons!
I have got from DRC mailing list a suggestion to play with octave scripts.
After some thinking I have decided to play with python as such funny games
potentially may be more usefull for me (probably it's my mistake :-)).
Thanks, I'll think what does windowing mean.
Happy aliking!
Andrew
2 Jul
2 Jul
12:58 p.m.
Being under DSP studying, I have accidentally met Judith C. Brown paper about
constant Q transform. An apparent thought is: constant Q transform is more direct
way for SPL-observing rather FFT+smoothing. Is it so (I don't mean RT-apps)?
======= On Saturday 24 June 2006 19:30, Fons Adriaensen wrote: =======
On Sat, Jun 24, 2006 at 07:15:38PM +0400, Andrew Gaydenko wrote:
And the question is: what is common way to smooth the
result? Some offtopic apps has
something like "1/24, ..., 1/3 octave smoothing". What does it mean?
Lowpass (linear phase) filtering of the response. Or windowing the
IR which is equivalent.
If you window the IR, you get a constant resolution over the
entire range of the FR. To get a variable rate such as 1/3 oct,
you need frequency dependent windowing - the window gets shorter
as frequency rises. It's not easy to do right (one of the reasons
why aliki still isn't finished).
IIRC DRC can produce such plots if you ask it nicely.
--
FA
Follie! Follie! Delirio vano e' questo!
3 Jul
3 Jul
5:08 a.m.
Hello Andrew,
On Sun, July 2, 2006 16:57, Andrew Gaydenko wrote:
...
way for SPL-observing rather FFT+smoothing. Is it so
(I don't mean RT-apps)?
usually, yes. AFAIK there's no transform that really mimimcs our perception,
i.e. which provides a visual representation similar enough to the auditory
perception. Take also a look at the Stockwell transform, which is a somewhat
balanced compromise between the properties of the FFT and the properties of
the wavelet transform.
...
IIRC DRC can produce such plots if you ask it nicely.
BTW frequency dependent windowing and standard fractional octave smoothing
aren't completely equivalent. There's instead a proven equivalence between
frequency dependent windowing and *complex* smoothing. See the Mourjopoulos
papers for further details.
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
5 Jul
5 Jul
11:02 p.m.
======= On Monday 03 July 2006 11:07, Denis Sbragion wrote: =======
Hello Andrew,
...
See the Mourjopoulos
papers for further details.
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
======================================================================================
Denis,
I have not found free description of the method defined in this article:
P.Hatziantoniou,J.Mourjopoulos,"Generalized Fractional-Octave Smoothing of Audio and
Acoustic
Responses", Journal of the Audio Engineering Society, Vol. 48, No. 4, pp. 259 - 280,
April 2000.
At any case, I have tried very simple smoothing of FFT-output, when y(n) is an arithmetic
mean
of x(n) and few other input signals before x(n). This "few other" is
proportional to n. Result
of "1/6 octave smoothing" is here:
http://gaydenko.com/mix/simpleSmoothing.png
"For eyes", the result is absolutely appropriate for my audio-DIY-ing purposes.
Probably, shown below short python fragment is better rather my ugly English. The
fragment
is applied just after FFT-ing (i.e., before converting to db).
Andrew
===============================
import Numeric
def smooth( x, # real
octaveFraction = 1.01 # for "1/3 octave smoothing"
# this par is supposed to be a pow(2.0, 1.0/3.0)
):
y = []
for n in range(len(x)):
sum = x[n]
meanLength = int( (octaveFraction - 1.0) * n / 2.0 )
if meanLength > 0:
for k in range(meanLength):
idx = n - k
if(idx < 0): idx = 0 # to avoid edge effect
sum += x[idx]
y.append(sum / (meanLength + 1))
return Numeric.array(y)
6 Jul
6 Jul
5:19 a.m.
Hello Andrew,
On Thu, July 6, 2006 03:02, Andrew Gaydenko wrote:
At any case, I have tried very simple smoothing of
FFT-output, when y(n) is an
arithmetic mean
of x(n) and few other input signals before x(n). This "few other" is
proportional to n. Result
of "1/6 octave smoothing" is here:
http://gaydenko.com/mix/simpleSmoothing.png
...
this looks pretty much like the classical "running average" smoothing, which
is one of the standard methods used to perform fractional octave smoothing.
For most purposes this is pretty good. BTW there's one thing that is unclear
to me:
...
if(idx < 0): idx = 0 # to avoid
edge effect
sum += x[idx]
...
Here x[idx] is a complex value or a real one (i.e. magnitude only)? If it's a
real one you're doing the standard fractional octave smoothing, if it is a
complex one you're doing something pretty close to complex smoothing.
Looking at the results it should be a real value, else there should be the
typical phase cancellation problems of the complex smoothing procedure. There
are some way to overcome these problems. I have a paper where smothing is
performed separately on the magnitude and phase of the response. The results
are really good in this situation, as long as you are able to compute the
continuous (unwrapped) phase function, which is a challenging task by itself.
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
10:03 a.m.
======= On Thursday 06 July 2006 11:19, Denis Sbragion wrote: =======
Hello Andrew,
...
this looks pretty much like the classical "running average" smoothing, which
is one of the standard methods used to perform fractional octave smoothing.
For most purposes this is pretty good. BTW there's one thing that is unclear
to me:
...
if(idx < 0): idx = 0 # to avoid
edge effect
sum += x[idx]
...
Here x[idx] is a complex value or a real one (i.e. magnitude only)? If it's a
real one you're doing the standard fractional octave smoothing, if it is a
complex one you're doing something pretty close to complex smoothing.
Looking at the results it should be a real value, else there should be the
typical phase cancellation problems of the complex smoothing procedure. There
are some way to overcome these problems. I have a paper where smothing is
performed separately on the magnitude and phase of the response. The results
are really good in this situation, as long as you are able to compute the
continuous (unwrapped) phase function, which is a challenging task by itself.
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
=================================================================================
Denis,
x[n] is real (abs() of the FFT result). It is interesting, I still have not found any
article about fractional octave smoothing. All googling results lead to cited AES
article.
So, applied smoothing is as an "invention" for me :-)
Is your paper about smoothing free?
Andrew
10:57 a.m.
Hello Andrew,
On Thu, July 6, 2006 14:03, Andrew Gaydenko wrote:
x[n] is real (abs() of the FFT result). It is
interesting, I still have not
found any
article about fractional octave smoothing. All googling results lead to cited
AES article.
So, applied smoothing is as an "invention" for me :-)
I suspect it is an invention for almost everyone... :) There's little
documentation about the common smoothing procedures used in audio. Most of the
available papers have been written after the Mourjopoulos ones. A better
analysis of the effects of smoothing has been started because an accurate
smoothing procedure of the complex response is needed for digital room
correction purposes. Before this it was pretty much just trickery used to make
the results easier to read to our eyes.
Is your paper about smoothing free?
"The use of continuous phase for interpolation, smoothing and forming mean
values of complex frequency response curves"
It is available here:
http://www.randteam.de/papers_jp/aes116_The_use_of_continuous_phase-JP.pdf
Bye,
--
Denis Sbragion
InfoTecna
Tel: +39 0362 805396, Fax: +39 0362 805404
URL: http://www.infotecna.it
6512
days inactive
6524
days old
linux-audio-dev@lists.linuxaudio.org
8 comments
3 participants
participants (3)
-
Andrew Gaydenko -
Denis Sbragion -
Fons Adriaensen